Microsoft Patents a System for Compressing and Unifying Mixed Medical Data
Hospitals generate dozens of different kinds of data, from MRI images to medication logs, and getting them to talk to each other is a genuine mess. Microsoft has filed a patent for a system that compresses each data type separately, then maps them all into a shared format so they can be compared and analyzed together.
How Microsoft wants to tidy up messy hospital data
Today, hospitals store patient information in wildly different formats. An MRI scan, a blood test result, and a doctor's written note all live in separate systems that were never designed to be combined, which makes it hard for software to analyze a patient's full picture at once. Microsoft wants to change how that data gets organized before it even reaches an AI or analytics tool.
The system described in this patent takes each type of medical data, compresses it to save storage space, then converts it into a common format. Once everything speaks the same language, records from different sources can be linked together by patient, time, or condition.
For you as a patient, this kind of behind-the-scenes plumbing is what makes it possible for a hospital's software to pull your complete history rather than just one silo of records. It is not a glamorous idea, but getting it right is what separates a useful health AI from one that only sees part of the picture.
… transforming the first modality data into normalized first modality data by: transforming the first delta table of the first schema into a first delta table of a second schema …
Translation: It converts the compressed data into a standard format so different types of medical records can be compared.
Inside Microsoft's two-stage compression and schema pipeline
The patent describes a two-stage pipeline for handling what it calls multimodal healthcare data, meaning data that comes in fundamentally different forms, such as images, sensor readings, structured lab results, and free-text notes.
In the first stage, each data type (called a modality) is compressed into a delta table, which is a storage format that records only what changed rather than storing a full copy every time, saving significant space. This happens inside a first processing datastore, essentially a staging area.
In the second stage, each compressed delta table is transformed into a normalized schema (a standardized structure, the way a spreadsheet has consistent column names regardless of who filled it in) and moved to a second processing datastore. For data types like medical images that carry embedded metadata, the system extracts those tags separately before normalizing them.
Once all modalities share the same schema, the system associates them, linking a patient's image data to their lab data to their clinical notes in one place. Microsoft's steady investment in healthcare data infrastructure shows up across several filing areas, but this patent focuses specifically on the compression and normalization layer that sits before any analysis or AI processing happens.
The normalized first modality data is associated with the normalized second modality data in the second processing datastore.
Translation: The system links different types of standardized patient records together in a central storage location.
What this means for healthcare AI and patient records
Healthcare AI systems are only as good as the data they can see. If a model trying to predict a patient's risk of hospital readmission can only access typed notes but not lab values or imaging metadata, its predictions will be incomplete. This patent addresses the plumbing layer that determines what data actually reaches those models.
For health systems and the software vendors that serve them, a standardized compression-and-normalization layer could reduce storage costs and make it easier to run analysis across different departments or hospitals that historically used different record systems. That is a real operational problem, and a solved version of it would have meaningful downstream effects on both clinical software and patient care.
That makes this Microsoft's 391st filing in our Microsoft coverage since May, adding to work like the verified-sender security application and the AI training data router.
The core tradeoff here is between generality and fidelity. Normalizing every medical data type into a single shared schema is useful precisely because it makes data comparable, but that process almost always loses something. A schema designed to hold both an ECG waveform and a medication log has to abstract away the details that make each format meaningful, and the patent does not describe how it handles that information loss.
The delta-table approach to compression is a sensible choice for data that updates incrementally, like ongoing patient vitals, but it is less obviously suited to large static files like high-resolution scans. Whether the system gracefully handles that distinction, or treats all modalities the same way and accepts the inefficiency, is left unclear in the filing.
This reads as infrastructure work, the kind of thing that matters a lot operationally but does not in itself create a clinical capability. If the normalization layer is designed well, it becomes a useful foundation. If the schema choices are too rigid, it becomes another silo in disguise, just a tidier one.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
3 drawing sheets from US 2026/0269028 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →